Two models fit the same 15 data points. Model A has mean absolute error 4.8 on the training data and 5.1 on a 5-point holdout set. Model B has training error 2.1 but holdout error 12.6. Which model is better for future predictions?